On September 15, 2021, oneAPI Developer Summit came to an end with great success. The virtual conference was focused on oneAPI product suite , Data-Parallel C++ (DPC++) language and AI for accelerated computing across XPU architectures — CPU, GPU, FPGA and other accelerators.
Hosted by Intel Software in association with Analytics India Magazine, the event witnessed participation from more than 1000 attendees and 18+ speakers, who showcased their work on heterogeneous computing and artificial intelligence — for faster application performance, more productivity and greater innovation.
Some of the prominent industry leaders and academic speakers that participated in the event include Andrew Richards, Henry Gabb, Gustavo Acosta, Himanshu Shekhar, Sameer Shende, Swarnendu Biswas, Soumyadeep Bhattacharya, Sunil Sherlekar, Vinayaka Ram Gururajan, Shrirang Karandikar, Sanjay Wandhekar, Sachin Kelkar, Kavita Aroor, Chris Kachris, Senganal Thirunavukkarasu, and Amarpal Kapoor.
Delivering close to 18+ talks, the developer summit created a dialogue and generated content in the form of a keynote, tech talks, lightning talks, panel discussions, and hands-on workshops. From scientific simulations to the hardcore nitty-gritty of middleware implementations, the oneAPI DevSummit discussed various topics across the spectrum.
Kickstarting the developer event, Joe Curley, Senior Director of oneAPI products, solutions and ecosystem, Intel, spoke about some of the challenges in accelerating computing programmes and their plans of building a strong community of oneAPI users to scale innovations.
“oneAPI is our technology. It is built on two premises — one that you actually do need, an open standard, community-driven way to programme accelerators — where we have standards-based in language and interfaces. Another is to have common libraries and common library interfaces, where we support non-Intel platforms,” said Curley.
Further, he said that Intel is committed to the open-source ecosystem where developers can pick up the bits and innovate for their own inventions on top of the basic infrastructure. “Our objective here is to create a productive, performance and trustworthy way to compete to write accelerated computing programmes,” he added.
Giving a keynote presentation, Henry Gabb, senior principal engineer – SATG, Intel, delved deep into oneAPI software abstraction for heterogeneous computing. Gabb said that Data-Parallel C++ provides explicit parallel constructs and offload interfaces to support a broad range of accelerators. “oneAPI also provides libraries for compute- and data-intensive domains like deep learning, scientific computing, video analytics, and media processing,” he added.
Further, he said that it also defines the low-level hardware abstraction or interfaces that allow a language runtime system to utilise a hardware accelerator for its vendors effectively.
(Source: Henry Gabb , Intel | oneAPI DevSummit 2021)
Initiating the tech talk session at oneAPI DevSummit 2021, Gustavo Acosta and Himanshu Shekhar from TCS touched upon the ‘TCS IE Dynaport Smartpoint Project.’ Here, the duo spoke about the domain-specific challenges and how the oneAPI solution helped them overcome them. Besides this, they also highlighted some of the use cases, solution design and performance, alongside sharing some of the learnings and experiences they encountered while working with the oneAPI platform.
At the lightning talk 2 — ‘TAU Performance System & E4S‘ — Sameer Shende, the research associate professor and the director for the performance research laboratory at the University of Oregon, described some of the top-performing systems, scaling scientific software stack (E4S), and how it can be integrated with oneAPI. The session presented some of the advances in TAU to support oneAPI. It also explained how a user might use TAU to instrument the code without modifying the application binary.
In the following session on ‘Audio from gravitational wave simulations‘, Andrew Pastrello, a PhD student at UNSW Sydney and a nuclear engineer at the Australian Nuclear Science and Technology Organisation, showcased a simple interface to synthesise audio from gravitational waveforms produced by binary black hole inspiral-merger-ringdown (IME) simulations, using a modified version of the EOBNRv2HM (effective-one-body-numerical-relativity version 2 higher modes) approximant included in LALsuite.
Some of the key takeaways from this session are:
- Implementation of kernels in DPC++ is generally straightforward
- The biggest difficulties in this project were in understanding and modifying an existing code, which is a problem not specific to DPC++.
In the “Race Detection on unified shared memory (USM)‘ lightning talk, Dr Swarnendu Biswas, an assistant professor at the department of computer science and engineering at IIT Kanpur, spoke about the unintentional concurrency errors, alongside discussing programme analysis to detect programming errors in CPUs and GPUs, and even some databases. This programme analysis is said to help in enhancing the debugging toolchain for DPC++ programmes with oneAPI.
In the ‘lightning talk’ session on ‘Lattice Boltzmann Solvers in DPC++,‘ Soumyadeep Bhattacharya, an architect of HPC and simulations platforms at SankhyaSutra Labs, spoke on the application side of oneAPI, where they have been collaborating with a lot of people to leverage their standard C++ goals to DPC++, and target the upcoming high memory bandwidth architectures — which is becoming crucial for bandwidth bound scientific and engineering simulations.
In another interesting talk, Andrew Richards, the CEO and co-founder of Codeplay, spoke on the topic ‘Building an open, safe, accessible AI & HPC ecosystem.’ He said that the world of AI and high-performance computing (HPC) is dominated by closed, proprietary software models. To get high performance today, systems need accelerators that have high levels of parallelism. “However, with programming models like CUDA — how can we open this up? How can we make these models safe enough to drive a car? And how can we get industry to work together with industry standards?” questioned Richards.
Working on these challenges for years, Richard and his team at Codeplay have come up with multiple solutions to help build a safe solution accessible to the AI & HPC ecosystem with the help of OpenAI. In this session, he discussed some of the progress made, alongside sharing the roadmap for the open-source ecosystem.
One of the key highlights at oneAPI DevSummit was the panel discussion on ‘reskilling of developers for HPC and AI.’ The panel discussion was moderated by Sachin Kelkar, head of the Asia Pacific region – developer relations at Intel, alongside the panellists of industry experts like SankhyaSutra Labs’ CEO Dr Sunil Sherlekar; TCS’ Vinayaka Ram Gururajan; AlgoAsylum’s Shrirang Karandikar; and C-DAC’s Sanjay Wandhekar.
The panelist discussed some of the challenges and changes over the few years in terms of developer expectations on learning new toolsets, methodologies, and the role of communities. They also brought a unique perspective from HPC research, AI solutions, system integration fields, and more.
In the second half of the summit, the Developer Marketing Manager – APJ at Intel, Kavita Aroor, highlighted some developer ecosystem programs such as Intel Software Innovator and oneAPI Certified instructor program. In addition, her session explored DevCloud, oneAPI learning path modules and submission of DevMesh projects. Kavita also played an exciting Jeopardy game where the developers had the chance to show off their oneAPI and DPC++ knowledge and win some exciting prizes.
Check out DevMesh contest rules here.
In the last lightning talk, Chris Kachris, co-founder and CEO of InAccel, delivered interesting insights and an overview of ‘Accelerated ML on Intel® FPGAs using oneAPI‘ and showed how users could run ML applications up to 40x faster-using oneAPI.
Last but not least, Dr Senganal Thirunavukkarasu, a computational scientist (PhD) from North Carolina State University, and director of data science at Tiger Analytics, Singapore, spoke about — ‘use of Intel® oneAPI libraries like oneDNN, oneMKL to optimise the Tiger Demand Forecasting Solution,’ which used MXNET framework. In addition to this, he also touched upon some of the challenges faced during this project and how they overcame them to adopt Intel oneAPI libraries successfully.
Finally, ending the oneAPI DevSummit 2021 successfully, in a hands-on session on “Intel® oneAPI Toolkits for HPC and AI applications,” Dr Amarpal Kapoor and Lakshmi Narasimhan from Intel explained in detail the nuances of openAI, alongside revealing some of the interesting case studies like porting an HPC application to Intel GPUs and deploying an AI application on a traditional HPC cluster.
oneAPI in a Nutshell
oneAPI is a cross-industry, open-source programming model that delivers common developer experience across accelerator architectures. In addition, the platform encourages collaboration on the oneAPI specification and compatible oneAPI implementations across the ecosystem.
Its specification extends existing developer programming models to enable a diverse set of hardware through language, a set of library APIs, and a low-level hardware interface to support cross-architecture programming. Its oneAPI specification builds upon industry standards and provides an open, cross-platform developer stack to promote compatibility and enable productivity and innovation.
With your active Priority Support, you can get assistance through the Online Service Center for your Intel® Software Development Products.
- Submit questions, problems, and other technical support issues
- Monitor previously submitted issues